SILMA AI plugin for LiveKit Agents

Support for English and Arabic speech synthesis with SILMA AI TTS v2 — Modern Standard Arabic, the Saudi (Najdi) dialect, and English.

This plugin is maintained by SILMA AI and published independently of the LiveKit Agents repository.

Installation

pip install livekit-plugins-silma

Requires livekit-agents>=1.8.

Pre-requisites

You'll need an API key from SILMA. Set it as an environment variable:

export SILMA_API_KEY="..."

Usage

from livekit.agents import AgentSession
from livekit.plugins import silma

session = AgentSession(
    tts=silma.TTS(
        model="silma-tts-v2-msa",
        voice="omar",
    ),
    # ... stt, llm, vad
)

Models and voices

Cloned voices

Upload a voice under Custom Voices at https://app.silma.ai/voices and pass its id along with your user id:

silma.TTS(
    model="silma-tts-v2-ksa",
    voice="omar",
    user_id="...",
    custom_audio_id="voice_1769817467123",
)

Pronunciation hints

SILMA reads phone numbers, emails and links correctly when they are tagged in the text:

await session.say(
    "You can reach us on <STAG_PN>92005455</STAG_PN> or at <STAG_EMAIL>hi@silma.ai</STAG_EMAIL>."
)

The plugin keeps these tags intact when it splits text, so a tag is never cut in half across two requests.

Account-level pronunciation overrides configured at https://app.silma.ai/control are applied when you pass user_id and enable_server_pronunciation_overrides=True.

How it works

stream() uses the realtime WebSocket API (wss://api.silma.ai/tts/v2/ws/stream) and sentence-tokenizes incoming LLM text so each request is a complete utterance. synthesize() uses the binary streaming endpoint (POST /stream).

SILMA caps text at 250 characters per request, so longer input is split on word boundaries and sent as sequential requests concatenated into one audio segment.

SILMA returns a 24 kHz mono float32 waveform; the plugin converts it to 16-bit PCM for the agent pipeline.

Development

pip install -e ".[dev]"
pytest

License

Apache-2.0